This report explains how DCTCP introduces 1--2 rounds of unnecessary lag, due to the way it processes congestion feedback. To solve this, a per-ACK moving average is proposed. It always cuts out 1 RTT of this lag and, paradoxically, it cuts out most of the rest of the lag by spreading the congestion response over a round. The EWMA still averages out variations in the feedback signal over the same set number of round trips, even though it is clocked per-ACK. This version of the report is released prior to full evaluation, in order to elicit early feedback on the design.
翻译:本报告解释了DCTCP如何因处理拥堵反馈的方式而引入1-2轮不必要的时滞。为了解决这个问题,建议了每包包平均移动平均值。它总是将这一时滞的1列多夫特切除,而且矛盾的是,它通过将拥堵反应分散在一回合中来削减大部分的时滞。EWMA仍然平均显示反馈信号在相同次数的往返旅行中的变异,尽管每轮旅行的时钟被锁定。报告的这一版本在全面评价之前发布,以获得对设计的早期反馈。